Scan the codebase and (re)build the structural graph using tree-sitter AST parsing. Auto-detects 20 languages and frameworks. Runs cross-stack resolvers (HTTP, gRPC, GraphQL, WebSocket, queues, events, CLI). Incremental by default — reuses a per-file parse cache so only changed files re-parse. Accepts a local path or a git URL (cloned on demand). Call on first use or after major refactors.

Repo overview: node/edge counts, detected kinds, entry points, and a dense text preview. Call this first to orient before using other tools.

Structural graph in dense sigil notation — the map of entities, relationships, and scopes. The primary context tool: feed it to the LLM so it can navigate without reading files. With seed, returns just that node's activated neighbourhood (scoped map) instead of the whole graph; otherwise the full map, truncated to budget.

End-to-end flow for a feature: entry point through service layer to data store, rendered as layered tiers. Call after status to drill into a specific feature.

Shortest path between two nodes, showing each hop with tier transitions.

Blast radius analysis: fan out from one or more nodes to see everything they affect (downstream) or depend on (upstream), grouped by tier. Pass several comma-separated nodes to assess a whole diff at once.

All direct connections to and from a node, one hop in each direction.

Return the source code for a node, sliced from its file by the graph's line span. Use after locate/find/activate to read the exact code without grepping. Returns a code block headed by the qname and path:start-end.

Spreading activation from seed nodes — finds the most relevant nodes in the graph relative to your seeds. Uses Personalized PageRank with domain-tuned edge weights. profile retunes the weights for a task (repair/review/onboard). mode=prose returns the ranked subgraph as primed prose instead of a score table.

Find nodes by name or qualified name pattern. Returns matching nodes with their kinds and qnames.

Resolve a failure signal — a stacktrace, a failing-test id, or a diff/changed-file list — to the most relevant nodes in the graph. Sniffs the signal shape, maps frames/symbols/paths to seed nodes, then ranks the surrounding subgraph by Personalized PageRank. The on-ramp for debugging: paste the error, get the code that matters.

Visual ASCII graph. With node: tree of children and connections. Without: full overview with counts.

Re-generate the graph from source. Incremental by default — only changed files re-parse, so this is cheap to call after edits. Set incremental=False to force a full reparse.